/ Machine Learning

Code of our winning entry to NTIRE 2018 super-resolution challenge

Code of our winning entry to NTIRE 2018 super-resolution challenge

wdsr_ntire2018

Wide Activation for Efficient and Accurate Image Super-Resolution.

Run

  1. Requirements:
    • Install PyTorch (tested on release 0.4.0 and 0.4.1).
    • Clone EDSR-Pytorch as backbone training framework.
  2. Training and Validation:
    • Copy wdsr_a.py, wdsr_b.py into EDSR-PyTorch/src/model/.
    • Modify EDSR-PyTorch/src/option.py and EDSR-PyTorch/src/demo.sh to support --n_feats, --block_feats option.
    • Launch training with EDSR-Pytorch as backbone training framework.

Overall Performance

Network Parameters DIV2K (val) PSNR
EDSR Baseline 1,372,318 34.61
WDSR Baseline 1,190,100 34.77

We measured PSNR using DIV2K 0801 ~ 0900 (trained on 0000 ~ 0800) on RGB channels without self-ensemble which is identical to EDSR baseline model settings. Both baseline models have 16 residual blocks.

More results:

Number of Residual Blocks13
SR NetworkEDSRWDSR-AWDSR-BEDSRWDSR-AWDSR-B
Parameters2.6M0.8M0.8M4.1M2.3M2.3M
DIV2K (val) PSNR33.21033.32333.43434.04334.16334.205
Number of Residual Blocks58
SR NetworkEDSRWDSR-AWDSR-BEDSRWDSR-AWDSR-B
Parameters5.6M3.7M3.7M7.8M6.0M6.0M
DIV2K (val) PSNR34.28434.38834.40934.45734.54134.536

Comparisons of EDSR and our proposed WDSR-A, WDSR-B for image bicubic x2 super-resolution on DIV2K dataset.

WDSR Network Architecture

Left: vanilla residual block in EDSR. Middle: wide activation. Right: wider activation with linear low-rank convolution. The proposed wide activation WDSR-A, WDSR-B have similar merits with MobileNet V2 but different architectures and much better PSNR.

Weight Normalization vs. Batch Normalization and No Normalization

Training loss and validation PSNR with weight normalization, batch normalization or no normalization. Training with weight normalization has faster convergence and better accuracy.

GitHub